CN106803301A - A kind of recognition of face guard method and system based on deep learning - Google Patents

Abstract

The invention discloses a face recognition access control method and system based on deep learning, wherein the method includes: receiving an access control removal request, judging whether the access control removal request is sent by a designated terminal, and if not, determining the access control removal request Sent by the image acquisition terminal; acquiring image information corresponding to the access control elimination request collected by the image acquisition terminal, performing face recognition on the image information using a deep learning face recognition algorithm, and obtaining a corresponding face recognition result; Judging whether the face recognition result corresponds to a preset face, if yes, instructs the access control system to remove the access control, and if not, refuses to remove the access control. Since a face can uniquely identify a person and is not easy to forge, in this application, the control of the access control system is realized through face recognition, which greatly improves the security of the access control system.

Description

A face recognition access control method and system based on deep learning

technical field

The present invention relates to the technical field of electronic equipment, and more specifically, to a face recognition access control method and system based on deep learning.

Background technique

With the improvement of people's living standards, people pay more attention to the safety of the home environment, and the concept of security continues to strengthen; with the improvement of this demand, intelligent access control systems have emerged, and more and more enterprises, shops, and families have installed various Various access control systems.

The currently more commonly used access control systems are nothing more than video access control, password access control, radio frequency access control or fingerprint access control and so on. Among them, video access control simply transmits video information to users without much intelligence. Forget it, and it is easy to crack; the disadvantage of radio frequency access control is "recognize the card but not the person", and the radio frequency card is easy to be lost and stolen by others; in addition, the security risk of fingerprint access control is that fingerprints are easy to copy. Therefore, the above-mentioned access control systems provided in the prior art all have the problem of low security for corresponding reasons.

To sum up, how to provide a corresponding technical solution for an access control system with higher security is an urgent problem to be solved by those skilled in the art.

Contents of the invention

The purpose of the present invention is to provide a face recognition access control method and system based on deep learning, so as to ensure the high security of the access control system.

In order to achieve the above object, the present invention provides the following technical solutions:

A face recognition access control method based on deep learning, comprising:

Receiving the access control elimination request, judging whether the access control elimination request is sent by a designated terminal, if not, then determining that the access control elimination request is sent by an image acquisition terminal;

Obtaining the image information corresponding to the access control removal request collected by the image acquisition terminal, performing face recognition on the image information using a deep learning face recognition algorithm, and obtaining a corresponding face recognition result;

Judging whether the face recognition result corresponds to a preset face, if yes, instructs the access control system to remove the access control, and if not, refuses to remove the access control.

Preferably, after obtaining the image information collected by the image collection terminal and corresponding to the access control elimination request, it further includes:

Use a deep learning photo recognition algorithm to identify the image information, and if it is recognized that the image information is obtained by shooting a real human face, perform the human face recognition algorithm for the image information using a deep learning face recognition algorithm. In the face recognition step, if it is recognized that the image information is obtained by taking a photo of a human face, then refuse to perform face recognition on the image information.

Preferably, it also includes:

If the face recognition result does not correspond to the preset face or the image information is obtained by taking a photo of a face, send an alarm message carrying the face recognition result or the image information to the Specifies the terminal.

Preferably, after sending the alarm information carrying the face recognition result or the image information to the designated terminal, it further includes:

Obtain the command information returned by the designated terminal after receiving the alarm information, execute the command information and store the command information and the corresponding face recognition result or image information, so that when the stored When the face recognition result or the image information is used, the corresponding command information is executed.

Preferably, it also includes:

If the face recognition result does not correspond to a preset face or is obtained by taking a photo of a face, a verification failure message is displayed to the outside world.

Preferably, it also includes:

If the access control cancellation request is sent by the specified terminal, instruct the access control system to cancel the access control.

Preferably, it also includes:

Use the infrared sensor of the human body to determine whether someone has entered the designated area, if yes, then instruct the image acquisition terminal to enter the normal working mode and collect image information, if not, then instruct the image acquisition terminal to keep the preset default sleep mode.

Preferably, after obtaining the image information collected by the image collection terminal and corresponding to the access control elimination request, it further includes:

The CCD image information and infrared image information included in the image information are fused, and the step of performing face recognition on the image information by using a deep learning face recognition algorithm is performed.

Preferably, using a deep learning face recognition algorithm to perform face recognition on the image information, including:

Face recognition is performed on the image information by using a deep learning face recognition algorithm based on GPU.

A face recognition access control system based on deep learning, including:

The first judging module is used to: receive the access control elimination request, judge whether the access control elimination request is sent by a specified terminal, if not, then determine that the access control elimination request is sent by an image acquisition terminal;

The image processing module is configured to: obtain the image information corresponding to the access control removal request collected by the image acquisition terminal, perform face recognition on the image information by using a deep learning face recognition algorithm, and obtain a corresponding face recognition result ;

The second judging module is used to: judge whether the face recognition result corresponds to a preset face, if yes, instruct the access control system to remove the access control, and if not, refuse to remove the access control.

The present invention provides a face recognition access control method and system based on deep learning, wherein the method includes: receiving an access control removal request, judging whether the access control removal request is sent by a designated terminal, and if not, determining the access control removal request Sent by the image acquisition terminal; acquiring image information corresponding to the access control elimination request collected by the image acquisition terminal, performing face recognition on the image information using a deep learning face recognition algorithm, and obtaining a corresponding face recognition result; Judging whether the face recognition result corresponds to a preset face, if yes, instructs the access control system to remove the access control, and if not, refuses to remove the access control. Among the technical features disclosed in this application, when the access control elimination request is sent by the image acquisition terminal, the image information corresponding to the access control elimination system is recognized by the deep learning face recognition algorithm, so as to determine whether the face recognition result of the image information is Corresponding to the preset face, if it is, it means that the owner of the face corresponding to the image information has the authority to remove the access control. At this time, instruct the access control system to remove the access control. Otherwise, it means that the owner of the face corresponding to the image information does not have the authority to remove the access control. At this time, it is refused to eliminate the access control. Since the face can uniquely identify a person and is not easy to forge, so in this application, the control of the access control system is realized through face recognition, which greatly improves the security of the access control system.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only It is an embodiment of the present invention, and those skilled in the art can also obtain other drawings according to the provided drawings without creative work.

Fig. 1 is a flow chart of a face recognition access control method based on deep learning provided by an embodiment of the present invention;

Fig. 2 is an algorithm structure diagram of a deep learning face recognition algorithm in a deep learning based face recognition access control method provided by an embodiment of the present invention;

3 is a model diagram of a deep learning face recognition algorithm in a deep learning-based face recognition access control method provided by an embodiment of the present invention;

Fig. 4 is a structure diagram of an image fusion algorithm in a face recognition access control method based on deep learning provided by an embodiment of the present invention;

5 is a flowchart of an image fusion algorithm in a face recognition access control method based on deep learning provided by an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a face recognition access control system based on deep learning provided by an embodiment of the present invention.

detailed description

The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Please refer to Fig. 1, which shows a flow chart of a face recognition access control method based on deep learning provided by an embodiment of the present invention, which may include:

S11: Receive the access control cancellation request, and determine whether the access control cancellation request is sent by a designated terminal, and if not, determine that the access control cancellation request is sent by the image acquisition terminal.

The designated terminal generally corresponds to the owner of the access control system, or the terminal of other pre-set people who have control authority over the access control system. If the access control elimination system is not sent by the designated terminal, it means that the image collection terminal has collected the images of the visitors. The information further sends an access control cancellation request. At this time, it is necessary to obtain the collected image information and determine whether the visitor corresponding to the image information can control the access control system.

S12: Obtain the image information corresponding to the access control removal request collected by the image acquisition terminal, perform face recognition on the image information using a deep learning face recognition algorithm, and obtain a corresponding face recognition result.

S13: Determine whether the face recognition result corresponds to a preset face, if yes, instruct the access control system to remove the access control, and if not, refuse to remove the access control.

Use the deep learning face recognition algorithm to perform face recognition on the image information. If the recognized face recognition result shows that the visitor corresponds to the preset face, it means that the visitor has the authority to control the access control system. At this time, the access control system is instructed to eliminate the access control. , that is to instruct the access control system to unlock the door and allow visitors to enter, otherwise, it means that the visitor does not have the authority to control the access control system, and refuses to remove the access control at this time. The preset face is the face information of the person who has control authority over the access control system.

Among the technical features disclosed in this application, when the access control elimination request is sent by the image acquisition terminal, the image information corresponding to the access control elimination system is recognized by the deep learning face recognition algorithm, so as to determine whether the face recognition result of the image information is Corresponding to the preset face, if it is, it means that the owner of the face corresponding to the image information has the authority to remove the access control. At this time, instruct the access control system to remove the access control. Otherwise, it means that the owner of the face corresponding to the image information does not have the authority to remove the access control. At this time, it is refused to eliminate the access control. Since the face can uniquely identify a person and is not easy to forge, so in this application, the control of the access control system is realized through face recognition, which greatly improves the security of the access control system.

In addition, the technical solution disclosed in this application adopts the most advanced face recognition algorithm based on deep learning to realize the recognition of image information. Compared with currently used algorithms such as PCA, SVM, and LBP, deep learning can accurately identify identity features. The recognition rate is higher, even surpassing the human naked eye recognition rate. Therefore, the technical solution disclosed in this application also has the characteristics of high face recognition accuracy, which further improves the security of the access control system.

In addition, it should be noted that the technical solution disclosed in this application focuses on face verification rather than face recognition, so that the difference between classes can be effectively reduced, it is easy to expand to other applications, and it is effective across databases; The more categories there are, the stronger the generalization ability is. Specifically, the field that this application focuses on is a subfield of face recognition—face verification, which simply means judging whether two pictures are the same person. In this way, the problem of face verification can be easily transformed into the problem of face recognition, which is to perform multiple face verifications. Use the deep learning method to learn a set of high-dimensional feature representations for face verification, and then learn the features by performing a multi-class classification face recognition task, and generalize the features to face verification and other unidentified faces. new identity verification. Identity features are taken from the activations of the last hidden layer. At the same time, multi-class classification is performed on all identities instead of two-class classification. This is based on two considerations: First, it is more difficult to train a training sample into one of multiple classes than to perform two-class classification. The challenge is to make full use of the super-learning ability of the neural network to extract effective features; the second is to implicitly add strong regularization to the convolutional neural network, which helps to generate shared hidden layer representations that are effective for classification. Therefore, the learned features have good generalization ability. The above algorithm architecture provided by the present invention can be shown in FIG. 2 .

The identification algorithm of this application is mainly composed of a deep convolutional neural network and a discriminative classifier, as shown in Figure 3, where the model parameters are as follows:

The first convolutional layer: the size of the convolution kernel is 4×4, and the number of channels is 3; the size of the output feature map is 36×36, with a total of 20 channels.

The first pooling layer: the kernel size is 2×2; the output sampling image size is 18×18, with a total of 20 channels.

The second convolutional layer: the size of the convolution kernel is 3×3, and the number of channels is 20; the size of the output feature map is 16×16, with a total of 40 channels.

The second pooling layer: the kernel size is 2×2; the output sampling image size is 8×8, with a total of 40 channels.

The third convolutional layer: the size of the convolution kernel is 3×3, and the number of channels is 40; the size of the output feature map is 6×6, with a total of 80 channels.

The third pooling layer: the kernel size is 2×2; the output sampling image size is 3×3, with a total of 80 channels.

The first fully connected layer: use the Maxout activation function to output a 160-dimensional vector.

The second fully connected layer: use the Maxout activation function to output a 160-dimensional vector.

The model inputs a 39×39×3 RGB three-channel color face image, and first passes through the first convolutional layer for feature extraction. The formula for extracting the feature map of the convolutional layer is:

f _ij = sigmoid((W*x) _ij +b)

The above formula means that the i row and j column pixels of the feature map are added by the convolution kernel and the convolution result of the same position of each channel of the input image to obtain the activation value. Wherein, W is the weight parameter of the neural network, b is the bias item parameter, and the activation function is sigmoid(z)=1/(1+e ^−z ). Compared with the full connection, the local convolution operation is easier to perceive local features, especially the facial features of the face, which can be extracted sensitively and can greatly reduce the weight parameters. However, there are still too many weight parameters, which are prone to overfitting and difficult to learn features. It is necessary to further reduce the parameters, so the pooling layer (also called sampling layer) is input after the convolutional layer. The pooling layer means that the local area in the feature map uses the same parameter, which can effectively reduce the parameters. The maximum pooling method is used here, and the sampling formula is as follows:

Where p _ij is the output image after pooling, x _ij is the input image, the purpose of the maximum value function in the above formula is to find the maximum pixel value point within the range of the pooling kernel, and the pixels in the same pooling area share the same weight parameter . This not only achieves lower-dimensional features, but also avoids the problem of overfitting.

Among them, the convolutional network has the following advantages over the general neural network in image processing:

· The topology of the input image and the network can be better matched.

· Feature extraction and pattern classification are performed simultaneously and are produced simultaneously during training.

· Weight sharing can reduce the training parameters of the network, making the neural network structure simpler and more adaptable.

Compared with the traditional convolutional neural network, the present invention uses a multi-layer perceptron + convolution to replace the original pure convolutional layer. Because convolution is a linear operation, it is not easy to learn nonlinear features, while multi-layer perceptrons have a strong ability to learn nonlinear functions. Based on this idea, a multi-layer perceptron is added before the original convolution layer to integrate the information between each channel to improve the generalization ability of the model. In practice, a multilayer perceptron can be implemented with a 1×1 convolution kernel. In addition, the Maxout activation function is used instead of the sigmoid function in the last two fully connected layers of this model. The reason is similar to that of the multilayer perceptron. The expression of the Maxout function is Maxout(x)=max(W ^T x+b). The Maxout function embodies the idea of function approximation. It uses continuous multi-slice linear functions to approximate nonlinear functions. The more slices, the better the approximation effect and the stronger the learning ability of nonlinear features. These two measures are to further improve the generalization ability of the model and enhance the model's ability to learn nonlinear features. The first step of the algorithm is to extract a high-dimensional face image feature. Thereafter the problem becomes a metric learning problem, where features are discriminated using a distance-based discriminative method, usually Euclidean distance. For images belonging to the same category (that is, the same person) in the training set, it is hoped that the Euclidean distance between similar training sets should be as small as possible; on the contrary, it is hoped that the Euclidean distance between training sets of different categories should be as large as possible. Based on this idea, the goal of defining a cost function is to let the convolutional neural network learn this cost function, thereby improving the generalization ability of the model as a whole.

Given an input image x, an image x p in the training set that belongs to the same class as x (that is, the same person, hereinafter referred to as the positive class), and an image _{x n} in the training set that is not in the same class as x (that is, not the same person, hereinafter referred to as the negative class) , f(x) represents the feature extracted by the image x through the convolutional neural network. First find two thresholds on the training set, and define the optimization objective function as:

The constraints are:

||f(x)-f(x _p )||+α＜||f(x)-f(x _n )||

where α is the maximum separation between positive and negative classes of f(x). The purpose of this optimization problem is to solve two thresholds a, b, and the optimal solution is make When ||f(x)-f(x _p )||＞a and ||f(x)-f(x _n )||＞b, it can be determined that x and x _p are the same person, otherwise they are not the same person . After solving the above problem optimization, we can get Then the cost function can be defined. The goal is to minimize the Euclidean distance between training sets of the same category and maximize the Euclidean distance between training sets of different categories. The problem can then be described as:

Putting the two problems together, we get:

So far, the cost function is obtained. In this way, the problem is transformed into an unconstrained convex optimization problem, which can be directly solved by stochastic gradient descent method or quasi-Newton method. Finally, the weight parameters of the convolutional neural network obtained by the stochastic gradient descent method or the quasi-Newton method are the optimal solution.

After extracting the features, it becomes a simple machine learning problem. Use SVM model for face verification. For binary classification problems, the SVM model has very good performance. The main strategy of the SVM method is margin maximization. Generally speaking, when classifying two sets of input spaces, it is always hoped to find a decision hyperplane that is far away from the two sets to distinguish them, because the distance between a point and the separating hyperplane can be Indicates the degree of confidence in the classification prediction, the farther away from the hyperplane, the more accurate the classification decision made. Based on this idea, the feature vectors that have been extracted from the convolutional neural network can be used as the training set to train an SVM model. Input two photos, extract two feature maps through the convolutional neural network, input these two feature maps into the SVM model, when the model outputs +1, it can be judged that the two pictures belong to the same person, and when the output is -1, it means two The pictures do not belong to the same person.

In addition, it should be noted that the door locks of the access control system in this application can use electric control locks. Specifically, the commonly used electric control locks in the access control system include electric mortise locks, magnetic locks, and electric locks. Among them, the electric mortise lock is mainly composed of two parts: the lock body and the lock hole, and the key part of the lock body is the "bolt". This type of electric lock drives the expansion and contraction of the "bolt" through the on and off of the current, and cooperates with the "magnet" to realize the function of locking or opening the door. It is precisely because of the retractable function of the "bolt" that it is called "electric mortise lock". In addition, its "hidden" installation feature is suitable for places that require high security of the lock body. The electromagnetic lock is an electronically controlled lock that relies on the suction force generated between the electromagnet and the iron block to close the door. It is a power-off door opening type. The usual model has a force of 280 kg. Due to the limited suction, it may be opened by many people or people with great strength. Therefore, electromagnetic locks are usually used in non-high-security places such as offices. If it is used in a safe place such as a prison, an electromagnetic lock with a tensile force of more than 500 kg must be customized. Therefore, the present application can use different types of electromagnetic locks according to different application occasions. The elimination of the access control system is to open the door lock, and the corresponding refusal to eliminate the access control system is to keep the closed state of the door lock.

The face recognition access control method based on deep learning provided by the embodiment of the present invention, after obtaining the image information corresponding to the access control elimination request collected by the image acquisition terminal, may further include:

Use the deep learning photo recognition algorithm to recognize the image information, if it is recognized that the image information is obtained by shooting a real face, then perform the step of using the deep learning face recognition algorithm to perform face recognition on the image information, if it is recognized If the image information is obtained by photographing the face of the photo, face recognition is refused for the image information.

It should be noted that before the image information is processed in this application, steps will be taken to prevent malicious spoofing of photos or video streams, and to confirm that the content of the photos taken is a photo of a real person rather than a photo of a face. Specifically, in real life, some criminals use photos or videos of valid users to attack the face recognition system. To solve this problem, some previous solutions are to prompt the visitor to make certain facial movements through voice (such as Blink, smile, etc.) to identify, to prevent criminals from stealing photos of valid users for malicious attacks. However, this method still has serious security risks: lawbreakers may also use valid user's facial high-definition video streams to attack the system. In addition, these similar methods need to add additional hardware devices, increase the cost of the system, and require users to make A certain posture and body coordination will greatly reduce the user's efficiency and experience. Based on these considerations, a non-intrusive real-time judgment method for real people and photos is provided only for a single photo source.

From the point of view of machine learning, this problem is the simplest classification problem - two-class classification, that is, to judge whether the content of a face photo is a real person or a photo. Let x be the input image, and y be the judgment result—assuming y=1 means that the input image is a real face, and y=0 means that the input image is a photo of a face. From the perspective of optical imaging, a real face has a three-dimensional structure, while a face photo has only a two-dimensional structure; compared with a real face, a face photo lacks one-dimensional information, and its reflected light should be relatively uniform, while a real face photo has only a two-dimensional structure. The reflected light of the face is random and belongs to diffuse reflection; the two images have different depth information. Using this mechanism and using the super learning ability of the deep convolutional network to extract the deep features of the image, the two images can be classified.

The architecture of the deep learning photo recognition algorithm disclosed in the present application can be shown in Figure 4, wherein the structure input layer, convolutional layer, pooling layer, convolutional layer, pooling layer, fully connected layer, Logistic regression layer. The convolution kernel size of the first convolution layer is 5×5 and the number of channels is 6; the convolution kernel size of the second convolution layer is 5×5 and the number of channels is 12; the window size of the two pooling layers is Both are 2×2. All activation functions used in this network architecture are sigmoid functions:

sigmoid(z)=1/(1+e ^-z )

The hypothetical function to be learned by defining this network is h _{W, b} (x), this function has a special probability meaning, it represents the probability that the output result is equal to 1, so the probabilities of the output result of the input image x being 1 and 0 are respectively:

P(y=1|x;W,b)=hW _,b (x)

P(y=0|x; W,b)=1-h _W,b (x)

The above two equations can be combined into one equation:

P(y|x; W, b) = (h _{W, b} (x)) y (1-h _{W, b} (x)) ^1-y y = 0, 1

Applying maximum likelihood estimation to this equation yields the loss function:

The second item is a regularization item, the purpose is to reduce the magnitude of the weight and prevent overfitting. During the training process, the convolutional neural network first calculates the error according to the above formula through forward propagation, and then calculates the partial derivative through error backpropagation, so that the parameters can be adjusted using the gradient descent method. The parameters when the final algorithm converges are the optimal optimal model. The above algorithm can directly judge a single static photo source. Compared with some dynamic target tracking methods, the judgment result is more reliable and has stronger real-time performance.

A face recognition access control method based on deep learning provided by an embodiment of the present invention may also include:

If the face recognition result does not correspond to the preset face or is obtained by photographing the face of the photo, an alarm message carrying the face recognition result or image information is sent to the designated terminal.

By sending the alarm information to the designated terminal, the designated terminal can perform corresponding operations according to the obtained information. If the corresponding user of the designated terminal determines that no one else is allowed to enter the access control system, the door lock of the access control system can be controlled through the designated terminal. Keep it closed, or if the visitor corresponding to the image information is the person who is allowed to eliminate the access control entry by the corresponding user of the designated terminal, you can control the opening of the door lock of the access control system through the designated terminal, so that the user of the designated terminal can realize the control of the access control system. remote monitoring, and then realize the corresponding control conveniently and quickly.

The face recognition access control method based on deep learning provided by the embodiment of the present invention, after sending the alarm information carrying the face recognition result or image information to the designated terminal, may also include:

Obtain the command information returned by the specified terminal after receiving the alarm information, execute the command information and store the command information and the corresponding face recognition result or image information, so as to execute the corresponding response when the stored face recognition result or image information is detected again command information.

The command information may include eliminating the access control system, keeping the door lock of the access control system closed, or automatically dialing 110, etc., which can be set according to actual needs, and are all within the protection scope of the present invention. Thus, after storing the corresponding information and the command information replied by the specified terminal user, the corresponding image information can be processed directly according to the stored command information in the later stage, and the control of the access control system is realized efficiently. In addition, the recognition events, alarm events, and all other events of any face device in the preset outlets can be sent to the background management center for real-time display and log records, which can be used as evidence in the event of a crime , and can be linked with the alarm system to push the alarm message to the APP to achieve a more three-dimensional security protection.

A face recognition access control method based on deep learning provided by an embodiment of the present invention may also include:

If the face recognition result does not correspond to the preset face or the image information is obtained by photographing the face in the photo, a verification failure message is displayed to the outside world.

The display module for the above display can be used by visitors to know their identity verification results. In addition, in order to facilitate interaction with users, the display module can use an LCD with a touch screen, which can make the operation of this work easier and convenient for users. You can also configure the network mode, set the status of the door lock, etc. through the LCD.

A face recognition access control method based on deep learning provided by an embodiment of the present invention may also include:

If the access control cancellation request is sent by the specified terminal, it instructs the access control system to cancel the access control.

The designated terminal with control authority to the access control system can be pre-set, so when it is determined that the access control removal request is sent by the designated terminal, it can directly instruct the access control system to remove the access control, so as to ensure the security of the access control system and facilitate the use of users.

A face recognition access control method based on deep learning provided by an embodiment of the present invention may also include:

Use the infrared sensor of the human body to judge whether someone has entered the designated area. If so, instruct the image acquisition terminal to enter the normal working mode and collect image information. If not, instruct the image acquisition terminal to maintain the preset default sleep mode.

The specified area can be set according to actual needs, such as within 3 meters from the door lock of the access control system. Since the entire access control system must work 24 hours a day, in consideration of power consumption, two working modes can be set for the access control system: normal working mode and default sleep mode. In the normal working mode, all the modules of the whole system are in the power-on working state, which consumes a lot of power, while in the sleep mode, only the human body infrared sensor is activated. When someone approaches the device, the infrared sensor module of the human body will sense that someone is approaching, and send a signal to request the processor to enter the normal working state and start all the modules. In general, the access control system defaults to sleep mode, and it will automatically enter sleep mode even after completing the corresponding operations after normal work, that is, when no one enters the preset area, it will remain in sleep mode.

Specifically, the human body infrared sensor is fully automatic. When a person enters its sensing range, it outputs a high level, and when the person leaves the sensing range, it automatically delays turning off the high level, outputs a low level, and the system receives this low level Then perform the corresponding wake-up operation. This module is used to reduce unnecessary resource consumption. If no one is present, the image acquisition process and network communication process in Linux do not sleep, and useless image data will be continuously collected and sent to the background for processing. This not only takes up resources in the background, but also increases the power consumption of this work.

The face recognition access control method based on deep learning provided by the embodiment of the present invention, after obtaining the image information corresponding to the access control elimination request collected by the image acquisition terminal, may further include:

The CCD image information and the infrared image information included in the image information are fused, and the step of performing face recognition on the image information by using a deep learning face recognition algorithm is performed.

It should be noted that CCD images and infrared images have their own advantages and disadvantages. In order to obtain clearer and more efficient image information, image fusion is performed on these two image information. Specifically, visible light images are most commonly used in daily life. For the human eye, visible light images have rich details and a keen sense of color, but it has poor penetration ability to the atmosphere under harsh weather conditions, and the imaging ability at night is also relatively poor; while infrared light is just right On the contrary, it has a strong penetrating ability under smoky environmental conditions. At night, due to the temperature difference between different objects, the image it forms can still show the outline of the object, but its disadvantage is the imaging resolution. lower. If the advantages of these two kinds of optical imaging are combined and the multi-spectral information is properly fused, the image blur caused by environmental factors can be eliminated, and then the target image with enhanced clarity can be obtained, and the detection and recognition capabilities of the target image can be improved.

Currently, the infrared and visible light fusion algorithms that are widely used are methods based on transform domains, such as wavelet transform, pyramid transform, and Contourlet transform. However, the above method does not have translation invariance, and it is easy to cause blurred image edge details. There is also a non-subsampling Contourlet transform (NCST) with translation invariance, but the algorithm complexity is too high. Because most of the existing algorithms are difficult to distinguish the characteristics of the noise and the original image, resulting in false or blurred information in the fused image. The invention uses an image fusion algorithm based on non-subsampling shearlet transform (NSST), which can greatly improve the efficiency of the algorithm. This algorithm first generates a saliency map from the infrared image, and then guides the infrared image to segment the target according to the saliency map, so that the infrared image with complex background or low signal-to-noise ratio can be accurately segmented. Then NSST transformation is performed on the infrared and visible light images respectively, and different fusion strategies are used for the target area (ie, the face area) and the background area of the two images. The main flow chart of this algorithm can be shown in Figure 5.

Among them, the infrared target area detection based on the saliency map involves salient target detection. Infrared imaging is related to the temperature of the object, so the target area (ie, the face) is prominent relative to the background area. Here, the salient region extraction method based on the frequency domain is used, and the Gaussian bandpass filter is selected to extract the salient features of the image. A Gaussian bandpass filter is defined as follows:

σ ₁ , σ ₂ (σ ₁ >σ ₂ ) are the standard deviations of the Gaussian filter, the low-frequency cutoff frequency is determined by σ1, and the high _- frequency cutoff frequency is determined by _σ2 . By choosing appropriate values of σ ₁ and σ ₂ , a saliency map that can maintain the expected spatial frequency characteristics can be obtained. The saliency map can be obtained by the following formula:

S(x, y)＝||I _μ -Iwhc(x, y)||

I _μ is the mean value vector of the infrared image, and I _whc (x, y) is the corresponding pixel value after Gaussian filtering. After obtaining the saliency map, according to the salient area in the saliency map, select the appropriate seed pixel points for image segmentation.

Target area fusion rules:

In order to preserve the thermal target information of the infrared image as much as possible, the low frequency subband coefficient of the infrared image is used as the low frequency band coefficient of the fusion image:

L ^F (x, y) = L ⁱ (x, y), (x, y)∈T

In order to strengthen the edge information, the high-frequency sub-band coefficients choose the "modulus maximum method".

The high-frequency sub-band and the low-frequency sub-band are all obtained through NSST transformation. L ^F , are the fused low-frequency sub-band and high-frequency sub-band coefficients, respectively.

Background area fusion rule: Use the fusion rule based on multi-resolution singular value decomposition to perform singular value decomposition on the matrix R:

R = USV ^T

Multiply U ^T by R to the left, and get A=U ^T R=SV ^T .

where S is a positive semi-definite diagonal singular value matrix. Arrange the singular values from large to small. The larger singular values correspond to the first few lines of A, corresponding to the low-frequency information in the image, which can represent the original appearance of the image to a greater extent, and the smaller singular values correspond to the next few lines of A, corresponding to high-frequency information. , which can reflect the image details. After that, rearrange the elements in the first few rows of A to obtain low-frequency subbands, and repeat the decomposition steps for the low-frequency subbands to achieve multi-resolution singular value decomposition.

A face recognition access control method based on deep learning provided by an embodiment of the present invention uses a deep learning face recognition algorithm to perform face recognition on image information, which may include:

Use the deep learning face recognition algorithm based on GPU to perform face recognition on the image information.

The present invention can be divided into three major parts from the structure: embedded equipment, background face recognition server, mobile client. Among them, embedded devices are developed based on ARM cortex-A series processors and equipped with Linux operating system. Because the chips of the cortex-A series have good multimedia processing capabilities, and the combination of high data throughput and high performance can well satisfy network processing applications, Linux is an operating system that supports multi-users, multi-tasking, multi-threading and multi-CPU. Has powerful network performance. Since the deep learning algorithm does not run directly on the embedded device (in fact, the calculation of the deep learning algorithm is very large, and running directly on the embedded device will greatly reduce the efficiency), so the main function of the embedded device is to drive the CCD and infrared camera The module collects images; and the background server is where the deep learning algorithm is actually run. At the same time, the background uses the GPU acceleration algorithm to run to improve the calculation speed. The recognition result can continue to be returned to the embedded device and the mobile client through the network communication protocol. In addition, the mobile client may refer to the designated terminal in this application.

In addition, it should be noted that the data transmission between different terminals in this application can be realized by using the SDIO-WIFI module. This network card conforms to the IEEE 802.11b/g standard, which can ensure the stable and efficient transmission of network data, and its data transmission rate can be Up to 54Mbps. This module uses the SDIO interface, which is much faster than the WIFI module with the SPI interface. Specifically, the SDIO bus is similar to the USB bus. The SDIO bus also has two ends, one end is the host (HOST) end, and the other end is the device end (DEVICE). The design of HOST-DEVICE is to simplify the design of DEVICE. All All communications are started by issuing commands from the HOST side. As long as the DEVICE side can parse the commands of the HOST, it can communicate with the HOST, and the SDIO HOST can connect multiple DEVICEs. Since the system needs to send network requests frequently and interact with the server background for data interaction, in order to prevent the program from entering an infinite waiting state due to network congestion during network transmission, the system adopts the idea of multi-threaded programming, and temporarily Create a thread object, and notify the main thread through the message mechanism after getting the return, let the main thread analyze the returned result and visualize the feedback to the user, and release the network request thread resources.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, as shown in Figure 6, which may include:

The first judging module 11 is used to: receive the access control elimination request, judge whether the access control elimination request is sent by a designated terminal, if not, then determine that the access control elimination request is sent by the image acquisition terminal;

The image processing module 12 is used to: obtain the image information corresponding to the access control elimination request collected by the image acquisition terminal, use the deep learning face recognition algorithm to perform face recognition on the image information, and obtain the corresponding face recognition result;

The second judging module 13 is used for: judging whether the face recognition result corresponds to a preset face, if yes, instructs the access control system to remove the access control, and if not, refuses to remove the access control.

A face recognition access control system based on deep learning provided by an embodiment of the present invention may also include:

The third judging module is used to: after obtaining the image information corresponding to the access control elimination request collected by the image acquisition terminal, use the deep learning photo recognition algorithm to identify the image information, if it is recognized that the image information is obtained by shooting a real face If it is, the step of performing face recognition on the image information using the deep learning face recognition algorithm is performed, and if the image information is recognized as being obtained by shooting the face of the photo, the face recognition on the image information is refused.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The alarm module is used for: if the face recognition result does not correspond to the preset face or the image information is obtained by shooting the face of the photo, then send the alarm information carrying the face recognition result or image information to the designated terminal.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The storage module is used to: after sending the alarm information carrying the face recognition result or image information to the specified terminal, obtain the command information returned by the specified terminal after receiving the alarm information, execute the command information and store the command information and the corresponding face recognition result or image information, so as to execute the corresponding command information when the stored face recognition result or image information is detected again.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The display module is used for: if the face recognition result does not correspond to the preset face or the image information is obtained by shooting the face of the photo, display the verification failure information to the outside world.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The access control elimination module is configured to: if the access control elimination request is sent by a designated terminal, instruct the access control system to eliminate the access control.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The fourth judging module is used to: use the infrared sensor of the human body to judge whether someone enters the designated area, if yes, then instruct the image acquisition terminal to enter the normal working mode and collect image information, if not, then instruct the image acquisition terminal to keep Set the default sleep mode.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, which may also include:

The fusion module is used to: after obtaining the image information corresponding to the access control elimination request collected by the image acquisition terminal, fuse the CCD image information and infrared image information contained in the image information, and perform the image information processing by using the deep learning face recognition algorithm. The steps of face recognition.

The embodiment of the present invention also provides a face recognition access control system based on deep learning, and the image processing module includes:

The image processing unit is used for: using a GPU-based deep learning face recognition algorithm to perform face recognition on the image information.

The embodiment of the present invention also provides a detailed description of the relevant parts in a face recognition access control system based on deep learning. Please refer to the detailed description of the corresponding parts in the face recognition access control method based on deep learning in the embodiment of the present invention. I won't repeat them here.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (10)

1. a kind of recognition of face guard method based on deep learning, it is characterised in that including：

Receive gate inhibition and eliminate request, judge that the gate inhibition eliminates whether request is sent by designated terminal, if it is not, then determining described Gate inhibition eliminates request and is sent by IMAQ terminal；

Obtain the collection of described image acquisition terminal eliminates the corresponding image information of request with the gate inhibition, using deep learning people Face recognizer carries out recognition of face to described image information, obtains corresponding face recognition result；

Judge the whether corresponding default face of the face recognition result, if it is, indicate gate control system to eliminate gate inhibition, if No, then refusal eliminates gate inhibition.

2. method according to claim 1, it is characterised in that obtaining the collection of described image acquisition terminal with the gate inhibition Eliminate after the corresponding image information of request, also include：

Described image information is identified using deep learning photo array algorithm, if identifying that described image information is right Real face is carried out shooting what is obtained, then perform described is carried out using deep learning face recognition algorithms to described image information The step of recognition of face, if identifying that described image information is that the face of photo is carried out shooting what is obtained, refuse to institute Stating image information carries out recognition of face.

3. method according to claim 2, it is characterised in that also include：

If it is that the face of photo is shot that the face recognition result does not correspond to default face or described image information Obtain, then send and carry the warning information of the face recognition result or described image information to the designated terminal.

4. method according to claim 3, it is characterised in that transmission carries the face recognition result or described image After the warning information of information to the designated terminal, also include：

Obtain the designated terminal and receive the command information returned after the warning information, perform the command information and by institute State command information and corresponding face recognition result or image information is stored, know with the face for detecting storage again Corresponding command information is performed when other result or described image information.

5. method according to claim 2, it is characterised in that also include：

If it is that the face of photo is shot that the face recognition result does not correspond to default face or described image information Obtain, then outwardly show the information of authentication failed.

6. method according to claim 2, it is characterised in that also include：

If the gate inhibition eliminates request and is sent by the designated terminal, it indicates that the gate control system eliminates gate inhibition.

7. method according to claim 1, it is characterised in that also include：

Judge whether that someone enters in designated area using human body infrared inductor, if it is, indicating described image collection eventually The collection of image information is held into normal mode of operation and carries out, if it is not, then indicating described image acquisition terminal to keep advance The acquiescence park mode of setting.

8. method according to claim 1, it is characterised in that obtaining the collection of described image acquisition terminal with the gate inhibition Eliminate after the corresponding image information of request, also include：

The ccd image information and Infrared Image Information that will be included in described image information are merged, and perform the utilization depth The step of study face recognition algorithms carry out recognition of face to described image information.

9. method according to claim 8, it is characterised in that believed described image using deep learning face recognition algorithms Breath carries out recognition of face, including：

Recognition of face is carried out to described image information using the deep learning face recognition algorithms realized based on GPU.

10. a kind of face recognition door control system based on deep learning, it is characterised in that including：

First judge module, is used for：Receive gate inhibition and eliminate request, judge that the gate inhibition eliminates whether request is sent out by designated terminal Send, if it is not, then determining that the gate inhibition eliminates request and sent by IMAQ terminal；

Image processing module, is used for：Obtain the collection of described image acquisition terminal eliminates the corresponding image of request with the gate inhibition Information, recognition of face is carried out using deep learning face recognition algorithms to described image information, obtains corresponding recognition of face knot Really；

Second judge module, is used for：Whether the corresponding default face of the face recognition result is judged, if it is, indicating gate inhibition System eliminates gate inhibition, if it is not, then refusal eliminates gate inhibition.